The control
What it can really operate
This is what exists today, with its limits. The phone is controlled through the Android accessibility service, which you have to enable by hand.
The phone
Open an app, tap an element by its text, type into the focused field and read the screen. That is enough to chain tasks, but there are no gestures or swipes: it is those four actions.
The home
Samsung TV over the network (on, off, volume, channel, HDMI, launching apps), air conditioning and lamps over infrared with codes learned from your remotes, and the robot vacuum.
The computers
From the chat you pick which machine runs the next thing, see its load in real time and can cancel whatever is running. The agent works inside the folders that machine allows, and only those.
The infrastructure
Branches and merge requests, DNS and domain email, new servers and domains. Anything that costs money always asks for explicit confirmation.